Mercurial > hg > camir-aes2014
comparison toolboxes/FullBNT-1.0.7/bnt/examples/dynamic/HHMM/Map/Old/mk_map_hhmm.m @ 0:e9a9cd732c1e tip
first hg version after svn
author | wolffd |
---|---|
date | Tue, 10 Feb 2015 15:05:51 +0000 |
parents | |
children |
comparison
equal
deleted
inserted
replaced
-1:000000000000 | 0:e9a9cd732c1e |
---|---|
1 function bnet = mk_map_hhmm(varargin) | |
2 | |
3 % p is the prob of a successful move (defines the reliability of motors) | |
4 p = 1; | |
5 num_obs_nodes = 1; | |
6 | |
7 for i=1:2:length(varargin) | |
8 switch varargin{i}, | |
9 case 'p', p = varargin{i+1}; | |
10 case 'numobs', num_obs_node = varargin{i+1}; | |
11 end | |
12 end | |
13 | |
14 | |
15 q = 1-p; | |
16 | |
17 % assign numbers to the nodes in topological order | |
18 U = 1; A = 2; C = 3; F = 4; O = 5; | |
19 | |
20 % create graph structure | |
21 | |
22 ss = 5; % slice size | |
23 intra = zeros(ss,ss); | |
24 intra(U,F)=1; | |
25 intra(A,[C F O])=1; | |
26 intra(C,[F O])=1; | |
27 | |
28 inter = zeros(ss,ss); | |
29 inter(U,[A C])=1; | |
30 inter(A,[A C])=1; | |
31 inter(F,[A C])=1; | |
32 inter(C,C)=1; | |
33 | |
34 % node sizes | |
35 ns = zeros(1,ss); | |
36 ns(U) = 2; % left/right | |
37 ns(A) = 2; | |
38 ns(C) = 3; | |
39 ns(F) = 2; | |
40 ns(O) = 5; % we will assign each state a unique symbol | |
41 l = 1; r = 2; % left/right | |
42 L = 1; R = 2; | |
43 | |
44 % Make the DBN | |
45 bnet = mk_dbn(intra, inter, ns, 'observed', O); | |
46 eclass = bnet.equiv_class; | |
47 | |
48 | |
49 | |
50 % Define CPDs for slice 1 | |
51 % We clamp all of them, i.e., do not try to learn them. | |
52 | |
53 % uniform probs over actions (the input could be chosen from a policy) | |
54 bnet.CPD{eclass(U,1)} = tabular_CPD(bnet, U, 'CPT', mk_stochastic(ones(ns(U),1)), ... | |
55 'adjustable', 0); | |
56 | |
57 % uniform probs over starting abstract state | |
58 bnet.CPD{eclass(A,1)} = tabular_CPD(bnet, A, 'CPT', mk_stochastic(ones(ns(A),1)), ... | |
59 'adjustable', 0); | |
60 | |
61 % Uniform probs over starting concrete state, modulo the fact | |
62 % that corridor 2 is only of length 2. | |
63 CPT = zeros(ns(A), ns(C)); % CPT(i,j) = P(C starts in j | A=i) | |
64 CPT(1, :) = [1/3 1/3 1/3]; | |
65 CPT(2, :) = [1/2 1/2 0]; | |
66 bnet.CPD{eclass(C,1)} = tabular_CPD(bnet, C, 'CPT', CPT, 'adjustable', 0); | |
67 | |
68 % Termination probs | |
69 CPT = zeros(ns(U), ns(A), ns(C), ns(F)); | |
70 CPT(r,1,1,:) = [1 0]; | |
71 CPT(r,1,2,:) = [1 0]; | |
72 CPT(r,1,3,:) = [q p]; | |
73 CPT(r,2,1,:) = [1 0]; | |
74 CPT(r,2,2,:) = [q p]; | |
75 CPT(l,1,1,:) = [q p]; | |
76 CPT(l,1,2,:) = [1 0]; | |
77 CPT(l,1,3,:) = [1 0]; | |
78 CPT(l,2,1,:) = [q p]; | |
79 CPT(l,2,2,:) = [1 0]; | |
80 | |
81 bnet.CPD{eclass(F,1)} = tabular_CPD(bnet, F, 'CPT', CPT); | |
82 | |
83 | |
84 % Assign each state a unique observation | |
85 CPT = zeros(ns(A), ns(C), ns(O)); | |
86 CPT(1,1,1)=1; | |
87 CPT(1,2,2)=1; | |
88 CPT(1,3,3)=1; | |
89 CPT(2,1,4)=1; | |
90 CPT(2,2,5)=1; | |
91 %CPT(2,3,:) undefined | |
92 | |
93 bnet.CPD{eclass(O,1)} = tabular_CPD(bnet, O, 'CPT', CPT); | |
94 | |
95 | |
96 % Define the CPDs for slice 2 | |
97 | |
98 % Abstract | |
99 | |
100 % Since the top level never resets, the starting distribution is irrelevant: | |
101 % A2 will be determined by sampling from transmat(A1,:). | |
102 % But the code requires we specify it anyway; we make it all 0s, a dummy value. | |
103 startprob = zeros(ns(U), ns(A)); | |
104 | |
105 transmat = zeros(ns(U), ns(A), ns(A)); | |
106 transmat(R,1,:) = [q p]; | |
107 transmat(R,2,:) = [0 1]; | |
108 transmat(L,1,:) = [1 0]; | |
109 transmat(L,2,:) = [p q]; | |
110 | |
111 % Qps are the parents we condition the parameters on, in this case just | |
112 % the past action. | |
113 bnet.CPD{eclass(A,2)} = hhmm2Q_CPD(bnet, A+ss, 'Fbelow', F, ... | |
114 'startprob', startprob, 'transprob', transmat); | |
115 | |
116 | |
117 | |
118 % Concrete | |
119 | |
120 transmat = zeros(ns(C), ns(U), ns(A), ns(C)); | |
121 transmat(1,r,1,:) = [q p 0.0]; | |
122 transmat(2,r,1,:) = [0.0 q p]; | |
123 transmat(3,r,1,:) = [0.0 0.0 1.0]; | |
124 transmat(1,r,2,:) = [q p 0.0]; | |
125 transmat(2,r,2,:) = [0.0 1.0 0.0]; | |
126 % | |
127 transmat(1,l,1,:) = [1.0 0.0 0.0]; | |
128 transmat(2,l,1,:) = [p q 0.0]; | |
129 transmat(3,l,1,:) = [0.0 p q]; | |
130 transmat(1,l,2,:) = [1.0 0.0 0.0]; | |
131 transmat(2,l,2,:) = [p q 0.0]; | |
132 | |
133 % Add a new dimension for A(t-1), by copying old vals, | |
134 % so the matrix is the same size as startprob | |
135 | |
136 | |
137 transmat = reshape(transmat, [ns(C) ns(U) ns(A) 1 ns(C)]); | |
138 transmat = repmat(transmat, [1 1 1 ns(A) 1]); | |
139 | |
140 % startprob(C(t-1), U(t-1), A(t-1), A(t), C(t)) | |
141 startprob = zeros(ns(C), ns(U), ns(A), ns(A), ns(C)); | |
142 startprob(1,L,1,1,:) = [1.0 0.0 0.0]; | |
143 startprob(3,R,1,2,:) = [1.0 0.0 0.0]; | |
144 startprob(3,R,1,1,:) = [0.0 0.0 1.0]; | |
145 % | |
146 startprob(1,L,2,1,:) = [0.0 0.0 010]; | |
147 startprob(2,L,2,1,:) = [1.0 0.0 0.0]; | |
148 startprob(2,R,2,2,:) = [0.0 1.0 0.0]; | |
149 | |
150 % want transmat(U,A,C,At,Ct), ie. in topo order | |
151 transmat = permute(transmat, [2 3 1 4 5]); | |
152 startprob = permute(startprob, [2 3 1 4 5]); | |
153 bnet.CPD{eclass(C,2)} = hhmm2Q_CPD(bnet, C+ss, 'Fself', F, ... | |
154 'startprob', startprob, 'transprob', transmat); | |
155 | |
156 |